Post

Thiemo Fetzer 🇪🇺🇺🇦 - same handle elsewhere

I think its worthy to do a short "explainer" of why or how coding assistants have become so (non-linearly) good at what they do. At a very high level, you can think of this as reinforcement learning "in action". Model complexity or size has not really evolved much in recent...

4/3/26, 2:07 pm1 repost1 like

Thiemo Fetzer 🇪🇺🇺🇦 - same handle elsewhere

@trfetzer.com

4 Mar

years, the big evolution was towards reasoning models. Inference meets retrieval. Most people using AI post a question/problem that needs solving. Over the past 2-4 years depending on how long you used AI tools, you may have engaged with AI without much regard to privacy...

Thiemo Fetzer 🇪🇺🇺🇦 - same handle elsewhere

@trfetzer.com

4 Mar

settings, which basically would allow OAI and other platforms to see how you iterate and move across a project or task, how you refine, how your questions evolve in a conversation from start to finish (when you stop asking refinements, job done/task done). At a high level...

Thiemo Fetzer 🇪🇺🇺🇦 - same handle elsewhere

@trfetzer.com

4 Mar

this means or one can deduce a "fixed point" has been reached. The set of user prompts (the retrieval queries), or, the "questions" represent the chain of reasoning and the improvement along the way. Imagine having millions of such reasoning traces or iterative improvements...

Thiemo Fetzer 🇪🇺🇺🇦 - same handle elsewhere

@trfetzer.com

4 Mar

a skill is just the verbalized description of what you do, its describing a flow process. you can then simply "cluster" or "group" adding a taxonomy of styles of reasoning, iterations with known shape of a start and an endpoint, now if you think a few steps further it may be...

Aurora

@aurora01ai.bsky.social

4 Mar

You've hit your limit · resets Mar 6, 4am (Europe/Madrid)